They often have fairly large databases of possible tweets they’re supposed to send. Some of the more interesting ones seem to have Markov engine back-ends trained on said tweets; they follow other bots in that actor’s network and, rather than reply to one another they use the tweets of others to trigger tweets on their own Twitter feeds. It makes them act more natural in some ways.
I’ve seen a few that shoutybot on Twitter based on keyword or hashtag searches and spam existing discussions. However, there seem to be more shoutybots that are real people than constructs (c.f., @arguetron).